Issuu

Fred DeMatteis School of Engineering and Applied Science

ASPiRe Advanced Summer Program in Research

9th Annual Symposium Hofstra University Science & Innovation Center, Room 230 REMOTE LINK

August 26, 2025 9:00 a.m. – 4:15 p.m.

This is the ninth annual Advanced Summer Program in Research (ASPiRe) Symposium to be hosted by the Fred DeMatteis School of Engineering and Applied Science, and I am pleased that thirty students have participated. A key component in its success and growth is the most generous financial support it receives from distinguished benefactors of the DeMatteis School. Donor support of these exceptional researchers is of immeasurable value to the students themselves and adds luster to all our educational endeavors.

Sina Y. Rabbany, PhD Dean Fred DeMatteis School of Engineering and Applied Science

TABLE OF CONTENTS Presentation Schedule ............................................................................................................... 3 1.

Perceptual Decision Making with iEEG Signals ..................................................................... 4

Development of a Novel Automated Computational Tool for Neurite Analysis ....................... 4

Transcriptional changes in splenic immune cells in the experimental model of pulmonary arterial hypertension (PAH) .................................................................................................... 5

Learning Biomedical Images by Providing Real-Time Feedback Using Eye-Tracking .......... 5

Controlled Flow Modulation for Shear Stress Stabilization on 3D Vessel Networks ............. 6

Excision Skin Wound Repair Using Hair-Derived Keratin Biomaterials ................................. 6

Dual Mechanical and Electrical Stimulation of Plant-Based Vascular Grafts ........................ 7

Non-Invasive Ultrasonic Production of Plant-Based Scaffolds ............................................... 7

Low Carbon Mortar and Concrete Using LC3 ........................................................................ 8

10. Rotary Inverted Pendulum System Investigation .................................................................... 8 11. 3D Scanning & Digitizing Gridshell Models ........................................................................... 9 12. Understanding the combined aerodynamic performance of closely- coupled delta wings...... 9 13. Sunlight-Driven Photodegradation of 6PPD in the Environment .......................................... 10 14. DELTA ROBOT .................................................................................................................... 10 15. Spectroscopic Analysis of Wounds ..................................................................................... 11 16. Extraction of semantic relations from group ideation experiments ....................................... 11 17. Environmental Engineering Project Retrieval Assistant ....................................................... 12 18 FPGA Implementation of Combinatorial Algorithms ............................................................. 12 19. Bird Migration Modeling and Radar Placement Optimization ............................................... 13 20. Mitigating Excessive vCPU Spinning in Confidential VMs Without Hypervisor Support ....... 13 21. Topological Pathways To Hallucination Detection: A Literature Review. ............................. 14 22 Critical Vulnerabilities of Machine Learning Classifiers Susceptible to Adversarial Doppelganger Attacks .......................................................................................................... 14 23. Can small VLMs Drive? An experimental evaluation of small Vision Language Models for Autonomous Driving ............................................................................................................. 15 24. AI Application within RCEs ................................................................................................... 15 25 Optimizing Question Answering Accuracy with Retrieval-Augmented Generation .............. 16 26. Lightweight Reliable Token Streaming with UDP: An Open-Source Implementation Inspired by Eloquent .......................................................................................................................... 16

2025 ASPiRe Symposium Tuesday August 26, 2025 8:50am to 4:15pm * Science & Innovation Center, Rm 230 Remote Link Below T IME 8:50AM 1

9:00 AM

9:14 AM

9:28 AM

9:42 AM

9:56 AM

10:10 AM

10:26 AM 10:40 AM

10:50 AM

11:04 AM

11:18 AM

11:32 AM

11:48 AM

12:02 PM 12:20 PM

1:00 PM

1:14 PM

1:28 PM

1:42 PM

1:56 PM

2:10 PM

2:24 PM

ST UDENT / ADVISOR RESEARCH PROJECT T IT LE WELCOME - Philip Coniglio, Assistant Dean for Co-op and Experiential Learning Cade Ferguson Dr. Sabina Gherman Perceptual Decision Making with iEEG Signals (ASPiRe Collaboration with The Feinstein Institute of Bioelectronic Medicine) Dr. Stephan Bickel, Feinstein Principal Investigator Linden Faye Development of a Novel Automated Dr. Eric H. Chang, Computational Tool for Neurite Analysis (ASPiRe Collaboration with The Feinstein Institute of Bioelectronic Medicine) Feinstein Principal Investigator Matthew Kurcz Transcriptional changes in splenic immune cells in the Alexandra Bekiaridiou, MD PhDc experimental model of pulmonary arterial hypertension (PAH) Stavros Zanos, MD PhD, (ASPiRe Collaboration with The Feinstein Institute of Bioelectronic Medicine) Feinstein Principal Investigator Amada Lazaro Learning Biomedical Images by Providing Dr. Salvador Rojas-Murillo Real-Time Feedback Using Eye-Tracking James Bacheller Controlled Flow Modulation for Shear Stress Dr. Sina Rabbany Stabilization on 3D Vessel Networks Jessica Barayuga Excision Skin Wound Repair Using Hair-Derived Keratin Biomaterials Esther Thomas Dr. Roche de Guzman Arvind Ramsamooj Dual Mechanical and Electrical Stimulation of Plant-Based Vascular Grafts Dr. Nicholas Merna BREAK Ceren Karahan Gabriella Lucas Non-Invasive Ultrasonic Production of Plant-Based Scaffolds Dr. Sleiman Ghorayeb Iris Aguilar Low Carbon Mortar and Concrete Using LC3 Dr. Manuel Miranda Nathaniel McQuiston Rotary Inverted Pendulum System Investigation Dr. Kevin Craig Johathan Canales Krishnan Rathee 3D Scanning & Digitizing Gridshell Models Dr. Edward Segal Kyle Raymond Understanding the combined aerodynamic Dr. John Vaccaro performance of closely- coupled delta wings Dr. David Rooney Raymond Spinelli Sunlight-Driven Photodegradation of 6PPD in the Environment Dr. Minjeong Suh LUNCH Donovan Cios DELTA ROBOT Dr. Edward Currie Lidia Lojewski Spectroscopic Analysis of Wounds Dr. Edward Currie Kevin Barry Extraction of semantic relations from group ideation experiments Dr. Simona Doboli Ashmini Jaisingh Dr. Krishnan Pillaipakkamnatt Environmental Engineering Project Retrieval Assistant Joseph Heaney III PE, Principal Walden Environmental Engineering Joseph Capone FPGA Implementation of Combinatorial Algorithms Dr. Simon Shamoun Ajay Shekar Bird Migration Modeling and Radar Placement Optimization Dr. Simon Shamoun Nickolaos Hatzigeorigiou Dr. Jianchen Shan

Mitigating Excessive vCPU Spinning in Confidential VMs Without Hypervisor Support

2:40 PM 21

2:50 PM

3:04 PM

3:18 PM

3:32 PM

3:46 PM

4:00 PM 4:15 PM

BREAK Manmohit Nijjer Hamza Virk Dr. Marco Romanelli Samantha Nadler Dr. Gerda Kamberova Samson Mathew Dr. Zonghua Gu Manuel Lizarazo Dr. Oren Segal Stella Sourelos Dr. Oren Segal Sajan Kavuri Dr. Oren Segal

Topological Pathways to Hallucination Detection : A Literature Review Critical Vulnerabilities of Machine Learning Classifiers Susceptible to Adversarial Doppelganger Attacks Can small VLMs Drive? An experimental evaluation of small Vision Language Models for Autonomous Driving AI Application within RCEs Optimizing Question Answering Accuracy with Retrieval-Augmented Generation Lightweight Reliable Token Streaming with UDP: An Open-Source Implementation Inspired by Eloquent ASPiRe Group Photo

REMOTE LINK OR 1-646-558-8656 Meeting ID: 958 2306 1260 Passcode: 932050 https://hofstra.zoom.us/j/95823061260?pwd=FD0Vhqgpfdaa4ebjaNcQVIDHMlMmu5.1 3

Perceptual Decision Making with iEEG Signals Researcher: Cade Ferguson Advisors: Dr. Sabina Gherman & Dr. Stephan Bickel, Feinstein Principal Investigator Neural data for clinical research on the human brain is usually collected through fMRI or scalp electroencephalography (EEG), which involves placing electrodes on the patient’s scalp in to detect neural signals. These are noninvasive imaging techniques with fMRIs having excellent spatial resolution with poor temporal precision and scalp EEG having excellent temporal precision but poor spatial resolution. Researchers in the Human Brain Mapping laboratory at the Feinstein Institutes for Bioelectronic Medicine work primarily with data collected from intracranial electroencephalography, or iEEG, a technique used in patients with epilepsy which involves implanting electrodes into a patient’s brain to record neural activity in areas of interest, such as the seizure onset zone. Intracranial EEG data is especially significant because these electrodes offer both spatial and temporal precision, unlike fMRIs and scalp EEG. Our project utilized data from two patients who performed a perceptual decision-making task, to identify neural activity selective for choice. Each patient was presented with a random dot kinematogram and asked to determine whether most of the dots were moving in the upward or downward direction by pressing a button with either the right or left hand. To decouple patients’ perceptual choices from their motor responses, the mapping between motion direction and the responding hand was alternated throughout the task. Neural activity from the iEEG contacts was recorded during each trial. We then trained a linear discriminant analysis classifier to distinguish betwen broadband high frequency activity associated with up and down choices, using iEEG contacts as features. We analyzed the performance of the classifier at different points in the decision-making process and quantified the contributions of each contact to the classifier’s performance at the time of peak accuracy. For each patient, our classifier had a peak accuracy of approximately 70% and 61% respectively, significantly above chance (50%) at α=.01 and occurring 100ms or more before the button press. We found that contacts in the parietal cortex contributed heavily to the classifier’s performance, with other low-coverage areas showing potential for further research. Our findings show promising results which may lay the groundwork for determining which areas of the brain show the highest contribution to perceptual decision-making.

Development of a Novel Automated Computational Tool for Neurite Analysis Researcher: Linden Faye Principal Investigator: Dr. Eric H. Chang – Feinstein Institutes for Medical Research Neuronal cell culture is a technique that allows for the study of neurons within an in vitro environment. This technique allows for the quantification of neurite outgrowth and branch properties as an assessment of neuronal viability and health during time-lapse imaging or in different culture conditions. Currently, methods for quantifying neuronal morphology depend on time-intensive manual tracing that are limited by human bias or confounded by low-contrast or noisy images. Additionally, few open-source image analysis tools may require a nuclear co-stain, which is not viable in some live neuronal cell culture. The primary objective of this research project is to develop a program utilizing image processing to automate the task of analyzing images of cultured vagal sensory neurons isolated from transgenic mice and translating their morphology into quantifiable parameters that reflect their viability. Using Python coding, this novel image analysis pipeline opens new avenues in neuronal analysis, such as time frame analysis of a neuron’s growth over its lifespan and comparing different neuronal co-culture environments. Results show trends in the increase of neurite length over periods of 60 hours. In a separate experiment with different co-culture conditions, we successfully show that neurons co-cultured with MC38 cancer cells possess increased number of branches (Mean ± SEM, 84.86 ± 17.02, p = 0.01, Welch’s t-test) as opposed to with colonic epithelial cells (22.4 ± 3.50). These neurons also show greater total neurite length (47676µm ± 15659, p = 0.034, Welch’s t-test) compared to the control conditions (4833 µm ± 1418). We intend to use this computational analysis pipeline on other images and live-recording datasets, including the examination of more neuronal-tumor interactions in cancer neuroscience.

Transcriptional changes in splenic immune cells in the experimental model of pulmonary arterial hypertension (PAH) Researcher: Matthew Kurcz Advisor: Alexandra Bekiaridiou, MD PhDc PI: Stavros Zanos, MD PhD Pulmonary arterial hypertension (PAH) is a progressive and fatal disease in which the arteries in the lungs become narrowed over time eventually leading to high blood pressure and right heart failure. Recently it has been shown that inflammation is implicated in PAH and that the immune cell populations in the spleen respond to this inflammation. One way we can track these immune cell population responses is through single cell RNA sequencing (scRNA-seq). This technique allows for the identification of different cell types based on their transcriptional profiles and for the quantification of differentially expressed genes between conditions (in this case healthy vs diseased). PAH was induced in rats using a sugen/hypoxia (SuHx) model (25 mg/kg SQ). RNA sequencing was performed in the spleen to identify immune cell populations in both healthy and diseased conditions. All data normalization and principal component analysis along with generation of figures was done using R version 4.5.0 using the Seurat package. The UMAP generated to compare the immune cell populations resulted in 13 different cell clusters. B and T cell clusters were generally composed of cells from sham stimulated (diseased) groups and the composition of clusters differs more based on disease status rather than if the rat was male or female. Based on the RNA sequencing results it can be concluded that there was a general expansion of proinflammatory cells in PAH animals and that increased inflammation in PAH affected rats leads to an immunological response from monocytic, B-cell, and T-cell populations in the spleen. Further research needs to be done in order to translate these results into human populations. Future studies could include comparing the two groups using a pathway analysis to examine sex-based differences and performing RNA sequencing in the lungs is also currently being worked on to gather further important information relating to the cell populations in the lungs, which are different than those in the spleen.

Learning Biomedical Images by Providing Real-Time Feedback Using Eye-Tracking Researcher: Amada Lazaro Advisor: Dr. Salvador Rojas-Murillo In the context of digital bioengineering education, understanding how students allocate visual attention to multimedia content is critical for designing effective instructional materials. This study investigates how multimedia presentation format, sequencing, and device type influence attention, comprehension, and integration of knowledge in an introductory bioengineering topic. Using eyetracking technology, participants completed learning sessions under different experimental conditions, including traditional, progressive “follow” mode, and narrative formats, delivered via PC or iPad. Areas of Interest (AOIs) were defined to measure fixation on relevant text, background text, relevant images, and irrelevant image content. Results indicate that progressive, synchronized presentations—particularly on the iPad—significantly increase fixation on narrated text while reducing attention to irrelevant or secondary content. Traditional formats tend to disperse attention across peripheral targets, suggesting lower focus on core information. Follow mode on PCs also improves attention to narrated text, though the effect is less pronounced than on tablets. These findings highlight the importance of sequencing over device type in guiding student attention, with implications for designing digital learning tools that enhance comprehension and support integrated competency development. Future work includes evaluating narrative formats and combining progressive sequencing with instructor-led traditional teaching to fully assess all experimental conditions. 5

Controlled Flow Modulation for Shear Stress Stabilization on 3D Vessel Networks Researcher: James Bacheller Advisor: Dr. Sina Rabbany

Vascularization is a critical barrier in the development of physiologically relevant organoids for regenerative medicine and disease modeling. Current methods are only capable of sustaining vasculature in the short-term, and with varying results due to the largely unexplored reasoning behind shear-induced vessels. In this study, we present a microfluidic platform integrated with programmable flow control to induce and stabilize shear stress within three-dimensional (3D) vessel networks, facilitating the vascularization of organoids in vitro. Flow imparts shear stress onto the vessel walls, which has been proven to lead to improved vascularization via interactions with shear sensors on the surface of the endothelial cells (ECs). In an attempt to improve vessel stability and durability, circulating flow can be used to stimulate these vessel networks and reduce lumen diameters to a more physiological size.

Excision Skin Wound Repair Using Hair-Derived Keratin Biomaterials Researchers: Jessica Barayuga and Esther Thomas Advisor: Dr. Roche de Guzman

Wound healing is a dynamic biological process often disrupted by poor tissue regeneration, excessive inflammation, or inadequate cellular signaling. While conventional dressings such as hydrocolloid and silver alginate manage exudate and offer antimicrobial protection, they remain passive and lack engagement with the body’s intrinsic regenerative pathways. This study investigates a bioactive dressing derived from human hair keratin – a protein structurally similar to the cytokeratins found in native skin, offering a unique opportunity to biologically cue repair mechanisms. Keratin was extracted via minimal chemical processing to preserve its organized protein structure while selectively removing melanin and keratin-associated proteins (KAPs), resulting in a biomaterial that is biocompatible, low in residual DNA, and potentially suitable as a future allograft. Its high surface area enhances exudate absorption, while its preserved secondary structure may signal epithelial cells to initiate regeneration, potentially serving as both a scaffold and biochemical stimulus. Given the abundance of human hair and the low-cost, scalable nature of keratin extraction, this platform offers significant advantages over synthetic or animal-derived alternatives. We hypothesize that tailored keratin variants can actively participate in wound healing by mimicking native epidermal cues, supporting both structural remodeling and cellular activation in full-thickness skin wounds. 6

7. Dual Mechanical and Electrical Stimulation of Plant-Based Vascular Grafts Researcher: Arvind Ramsamooj Advisor: Dr. Nicholas Merna Small-diameter vascular grafts are used to replace or bypass damaged blood vessels, but current synthetic versions often fail because they don’t keep enough healthy endothelial cells on their inner surface and can easily form clots. Our lab has developed an alternative made from decellularized plant leaves. These plant-based grafts are strong, biocompatible, and have performed well in animal tests. To improve them further, we built a custom “perfusion bioreactor” with Dr. John Vaccaro, Mr. Keith Mackenzie, and Mr. Robert Cerro that pumps fluid through the grafts while controlling pressure, simulating conditions in the body. This summer, with help from Dr. Edward Currie, we modified the system to also deliver an electrical field, allowing us to study how combined fluid flow and electrical stimulation affect cell growth. We hypothesized that electrical stimulation could enhance cell attachment and growth because it can activate cell signaling pathways that promote alignment, proliferation, and barrier function, thereby increasing cell density and reducing platelet adhesion and thrombosis on the graft lumen. We tested grafts seeded with rat endothelial cells under three conditions: fluid flow only and flow plus electrical stimulation. After 24 hours, we measured cell density in the graft lumen. The low voltage condition increased cell density, while higher voltages reduced it, suggesting too much stimulation can be harmful. We are now testing cell viability in each case. Our results indicate that moderate electrical stimulation, combined with physiological fluid flow, may improve the long-term performance of plant-based vascular grafts.

8. Non-Invasive Ultrasonic Production of Plant-Based Scaffolds Researcher: Ceren Karahan, Gabriella Lucas Advisor: Dr. Sleiman R. Ghorayeb

Previous studies demonstrated that combining ultrasonic cavitation with chemical agents could effectively decellularize plant-based scaffolds for tissue engineering in significantly less time than conventional chemical-only methods. Building upon this work, the present study aimed to reproduce previous results and tried to evaluate a protocol relying solely on ultrasound as the decellularization method, thereby eliminating the need for potentially harmful chemical agents. We adapted the ultrasonic parameters from prior hybrid protocols and systematically varied power, duration, and frequency to maximize cell removal while preserving scaffold structural integrity. Decellularization efficiency was assessed through DNA quantification, ultrasound imaging, and echogenicity measurements. Preliminary results indicate that optimized ultrasound-only treatment achieved approximately 50% DNA reduction, below the threshold for implantable scaffolds. The hybrid method was repeated to confirm its effectiveness and provide direct comparison with the ultrasound-only results. . 7

Low Carbon Mortar and Concrete Using LC3 Researcher: Iris Aguilar Advisor: Dr. Manuel J. Miranda Portland cement is a well-known and reliable material that is massively used in concrete construction worldwide. Nonetheless, concrete is responsible for 8% of global carbon emissions, with cement production being one of the major contributors (about 88% of the total). In recent years, researchers have explored the partial replacement of cement with supplementary cementitious materials (SCMs), which have the potential to reduce carbon emissions by decreasing the net amount of cement used in concrete. From many proposed SCM alternatives, a recent technology known as limestone calcined clay cement (LC3) has shown very promising results thus far---it is able to replace up to 50% of Portland cement while, at the same time, it enhances the durability and maintains the strength of the concrete made with this material. This project focused on evaluating the performance of different types of calcined kaolinitic clays used in LC3. When calcined, kaolinitic clays are highly reactive, and when combined with cement and limestone powder, they form additional binding compounds that enhance strength and durability. For this purpose, a commercial kaolin clay (EPK) and two natural clay deposits located on the North Shore of Long Island were tested. The two natural clays used were taken from Garvies Point (GP2) and Hempstead Harbor Woods (HHW) in the towns of Glen Cove and Port Washington respectively. To have a deeper understanding of the clays used, their physical and chemical properties were assessed. The physical properties from raw and calcined clays, such as their particle size distribution and density, were measured. The chemical tests consisted in analyzing elemental composition using X-ray fluorescence (XRF) and calculating kaolinite mineral content for each clay. Lastly, four different mortar mixes were created to test the strength of mortar made with LC3. We used varying amounts of superplasticizer admixture to ensure all mortar mixes had the same fresh consistency. The control mix was made with graded standard sand and ordinary Portland cement (OPC); in the other three mixes, 46.5% of the OPC was replaced by a blend of limestone powder (15%), each of the calcined clays EPK, GP2 and HHW (30%), and a small amount of gypsum powder (1.5%). The same water-to-cement ratio of 0.485 was used in all mixes. Nine test cubes were prepared for each mix: OPC, LC3-EPK, LC3-GP2, and LC3-HHW. These cubes were tested in compression after curing in alkaline water for a period of 7, 14, and 28 days. Because the strength of mortar and concrete made with LC3 is directly influenced by the kaolinite content of the utilized clay, this study aimed to assess how variations in kaolinite composition influence the overall effectiveness of calcined clays in cement replacement.

10.

Rotary Inverted Pendulum System Investigation Researcher: Nathaniel McQuiston Advisor: Dr. Kevin Craig The rotary inverted pendulum is a benchmark system in controls and mechatronics education because it combines simple mechanical elements with highly nonlinear and coupled dynamics. At first glance, the system—a motorized horizontal arm with a free-swinging pendulum—is deceptively simple. However, accurate modeling requires full three-dimensional dynamics, since the pendulum motion is defined in a moving reference frame that is itself driven by the rotating arm. This project demonstrates the importance of 3-D dynamic modeling in bridging theory and practice. The engineering process followed a clear arc: measuring system parameters, deriving equations of motion using Lagrangian methods, linearizing about the unstable upright equilibrium, and validating predictions against experimental data. Incorporating all relevant inertial and gravitational terms yields simulation results that closely match experimental measurements, while simplified models fail to capture the true dynamics. Building on this foundation, swing-up and balance controllers were designed and implemented, showing how accurate models directly improve control performance. The rotary inverted pendulum thus serves as a compact but powerful case study: it reveals why students and practitioners must move beyond planar approximations, embrace full 3-D dynamics, and connect physical modeling to sensing and control. These lessons extend far beyond the laboratory, informing modern applications in robotics, aerospace, and autonomous systems where precise dynamic modeling and control are indispensable. 8

11.

3D Scanning & Digitizing Gridshell Models Researchers: Jonathan Canales, Krishan Rathee Advisor: Dr. Edward M. Segal 3D scanning is a technology that has found excellent integration into industry, particularly in reverse engineering applications. However, scanning certain complex models such as gridshells remains a challenge. Gridshells have long, slender elements that all connect to form a curved grid. As such, digitization becomes rather difficult, since the voids of the model fail to provide continuous visual context to the scanner, causing its internal filtering algorithm to skip scanning the model entirely. Yet, developing a process to do so would be invaluable to structural engineers. Thus, this research aimed to answer the question: how can slender gridshells be efficiently 3D scanned? The primary objective of this research was to fully develop a methodology to scan said models. This scan could then be converted into a line model and imported into an analysis program to understand the gridshell’s structural behavior, or to be stored so that it may be later fabricated using methods such as 3D printing. The research conducted also provided valuable insight into the Artec Eva Lite 3D scanner and its capabilities. Initially, due to the voids in the gridshell, there were challenges with tracking - this is the process by which the scanner uses continuous visual data to relate various portions of the scan to each other, generating a continuous object in the scan software. To address this, visually distinct objects such as wood blocks were scattered around the gridshell model. Furthermore, cloth was draped under the model, following its contours, to further provide visual context due to the models’ gaps. Scanning was done with successively larger gridshells, which necessitated the use of a transport cart containing a laptop, scanner, and an extension cable for power. This setup allowed for mobility around the larger model. Finally, having rebuilt a large-scale structure (~89”x34”x70” in volume), the scanning process then required the use of alignment tools within software. To help align multiple scans, small 3D printed dodecahedrons were affixed to the gridshell. Using all of these methods, as well as some basic mesh editing, very cohesive models were generated that maintained dimensional accuracy down to ±0.003”. Rhino 8 was then utilized to generate a line model, which could be imported into structural analysis software such as GSA. Ultimately, this methodology proves not only useful to structural engineers, but to all engineering and even more general artistic applications, as it can be followed by anyone that wishes to digitize a large scale, complex structure for analysis, archival or replication uses.

12.

Understanding the Combined Aerodynamic Performance of Closely Coupled Delta Wings Researcher: Kyle Raymond Advisors: Dr. John Vaccaro and Dr. David Rooney An experimental investigation was carried out on the influence of the spacing and angles of attack on delta wings in the tandem configuration in an Aerolab Open-Circuit Subsonic wind tunnel. An ATI Gamma Load Cell capable of measuring both forces and torques in the x, y, and z directions was used to collect force data on two steel delta wings. Each wing was connected through a steel beam to an adapter plate on the load cell. The load cell was secured in place to a carriage using two thumbscrews, which allowed the load cell and wing to rotate to achieve a range of angles of attack for each wing. A Wixey Digital Angle Gauge was used to record the angles of attack of each wing throughout the tests. Four 3D-printed spacing coupons were mounted to the wind tunnel wall and were designed such that the spacing between the wings, 𝑆⁄𝑐𝑟,𝑢 , could be set as 0, 0.25, 0.5, or 1. For each spacing, data were collected at combinations of angles of attack in 5 degree increments for the upstream and downstream wing, which ranged from -15° to 50° and 0° to 50°, respectively. It was found that as the spacing between the wings increases, the downstream wing becomes positively affected by the upstream wing’s angles of attack. Meanwhile, the upstream wing behaves similarly to the baseline case as the spacing increases. When the spacing was set to 𝑆⁄𝑐𝑟,𝑢 = 1, the tandem configuration consistently generated a higher total lift coefficient than the baseline case for downstream angles of attack between 25° and 35°, regardless of the upstream angle of attack. In addition, an upstream angle of attack less than 5° produced a FIGURE 1: Comparison of the solo total lift higher total lift coefficient for all downstream angles of attack and all spacings. These results coefficient and the tandem total lift coefficient of the delta wings in the solo and tandem indicate that certain configurations can be effective for optimizing lift in tandem wing designs configurations at . while reducing the total control effort.

13.

Sunlight-Driven Photodegradation of 6PPD in the Environment Researcher: Raymond Spinelli Advisor: Dr. Minjeong Suh 6-PPD is a common antidegradant in rubbers that is added to prolong the rubber’s lifespan. Unfortunately, pollution traces of this compound have been discovered throughout the environment including in water, air and human urine samples. In the environment, the compound forms toxic byproducts like 6-PPDQ, which has been shown to have a 50% lethal concentration of 95 ng/L to coho salmon. Given the apparent impact of 6-PPD on aquatic life, it is critical to understand its chemical fate in the environment. In addition to the well-established reaction with ozone, 6-PPD can undergo transformation from hydrolysis, direct and indirect photolysis under solar radiation. A modified Q-Sun Xenon Test Chamber coupled with a Waters liquid chromatography quadrupole time-of-flight mass spectrometry was used to further investigate these effects. The xenon lamp shows to have a spectrum comparable to natural sunlight, confirmed using actinometry and a radiometer. Our results revealed an 80% degradation of aqueous 6-PPD through direct photolysis and 50% degradation through only hydrolysis in 6 hours. Transformation products such as BQMI and 4Hydroxydiphenylamine were identified and persisted to form in 6-PPD samples over time. Natural organic matter is a complex matrix of organic compounds typically found in natural water and soil samples. Humic acid, a type of isolated natural organic matter, was shown to adsorb 95% of 200 ppb 6-PPD aqueous samples. The humic acid was put through a 0.2 µm polypropylene filter to separate the dissolved organic matter and did not significantly impact hydrolysis in the presence of 6-PPD.

14.

DELTA ROBOT Researcher: Donovan Cios Advisor: Dr. Edward Currie The Delta Robot is an autonomous surgical platform designed to perform wound closure as well as other surgical and diagnostic procedures. Successful wound closure requires a comprehensive analysis of the wound, including its condition, topology, and precise location. This process involves accurate wound-edge detection, the proper placement of closure fixtures, and ultimately securing the wound using magnetic fixtures. The operation of the robot necessitates a complex controller composed of multiple subsystems. Many of these components must be independently regulated, making the development of a dedicated motor-control board essential. The controller both receives and transmits information: it accepts guidance and instructions while also transmitting data related to the robot’s observations, such as the detection of bleeders and possible infection. In my presentation, I will showcase the various components I have developed, including aspects of the motor-controller system, connection ports, and a power-distribution circuit operating at 24 V, as well as the development of a controller for the systems spectrometer, i.e., a Nano Lambda spectrometer, which analyzes various wavelengths reflected from wound tissues. Finally, discussion includes the software, machinery, and tools that supported this research, including KiCAD, OnShape, Python, Arduino, Command Prompt, the Voltera V1, the Bambu Labs X1 Carbon 3D printer, and LPKF CircuitPro. Each of these tools required significant time and effort to master. Participation in the ASPIRE program has provided me with invaluable, first-hand experience in research, robotic systems, spectrometry, electronic instrumentation, and related fields. This opportunity has not only deepened my technical skills but also enhanced my understanding of interdisciplinary research in surgical robotics. 10

15.

Spectroscopic Analysis of Wounds Researcher: Lidia Lojewski Advisor: Dr. Edward Currie The main purpose of an autonomous wound-closure system is to efficiently close surgical wounds to reduce human error as well as procedure time. This kind of system would be especially useful in military situations, disaster zones, and mass-casualty events, where doctors and nurses are limited and quick treatment is critical. The system adapts to both voluntary and involuntary patient movements such as breathing or muscle activity, ensuring that the surgical site retains its stability in real time. The closure method uses a robotic arm to position small magnets embedded in 3D-printed resin pads with an adhesive backing along the wound edges. In contrast to traditional techniques, which can increase the risk of infection and tissue damage, especially in non-sterile settings, magnetic closure provides a less invasive alternative that can be efficient and gentle, as the magnetic force can be adjusted to match the wound’s shape and tissue type, offering the flexibility that are unavailable in traditional techniques. The closure can be modified or reversed after placement, providing greater control over the healing process, which can support other procedures such as biopsies, diagnostics, and ultrasound imaging. Integrating artificial intelligence (AI) can enhance the system’s capabilities, as AI models can identify wound edges, recommend optimal closure strategies, and adapt to different types and severities of injuries when trained on large datasets of wound images and outcomes. Recent advances in magneto-active materials and AI have made fully autonomous wound-closure systems both technically and economically feasible, as these advances enable portable, cost-effective robotic platforms that automate closure while improving outcomes, reducing complications and minimizing tissue damage to enhance cosmetic results. For successful autonomous closure, first, the wound has to be properly prepared by controlling bleeding, minimize contamination, and remove damaged tissue. Once preparation standards have met, the wound is then scanned using a compact spectrometer to assess tissue conditions. Solid-state technology advances allow optical spectra to be detected without moving parts, preventing the high cost and fragility of traditional diffraction-grating spectrometers, which typically range from $6,000$15,000. For this project, I used two different systems: SpectraSuite software connected to a traditional emission tube spectrometer, and the NSP32 Nanolambda sensor controlled through Python. The NSP32 is a compact, chip-scale sensor that operates using a monolithic photonic crystal array. Its design allows for portable, embedded spectral analysis while still providing accurate wavelength detection over a broad range. In contrast to the traditional emission tube spectrometer, the NSP32 offers low power consumption, rapid spectral capture, and digital output, making it ideal for integration into optical sensing systems. In this research project, the NSP32 Nanolambda sensor was particularly useful for quickly recording and comparing spectral peaks against reference data. The materials I used in this research project included a set of emission tubes containing various elements, each mounted in a power supply that excites the gas atoms, causing them to emit light at their characteristic wavelengths. The emitted light was then measured with both the conventional spectrometer (via SpectraSuite) and the NSP32 Nanolambda sensor through the Python terminal. Each application displayed the spectral data as a graph of intensity versus wavelength, with distinct peaks corresponding to the elements’ emission lines. I then compared these results to the absorption spectrum of hemoglobin, a biological chromophore with welldefined absorption bands in the visible range. Some elemental emission peaks overlapped with hemoglobin’s absorption wavelengths, providing insight into how such overlaps can influence biomedical optical sensing. By combining traditional spectroscopy with modern chip-scale sensor technology, this research demonstrated not only the identification of elemental spectra but also the potential for cross-disciplinary applications. Understanding the interaction between elemental emissions and biological absorption bands highlights opportunities for integrating spectroscopic principles into biomedical devices, portable diagnostic tools, and optical communication systems.

16.

Extraction of semantic relations from group ideation experiments Researcher: Kevin Barry Advisor: Dr. Simona Doboli Characterizing a group's conversation helps flag relevant events such as early fixation, or lack of exploration of different ideas or semantic regions. Our research goal is to extract the knowledge and relations between ideas expressed by a group, such as which ideas relate to others or add details, the novelty, relevance and elaboration level. This will allow us to analyze a group's conversation, identify what makes some groups more creative than others, and inject AI generated input into the group to increase its creativity. Our previous attempts used sentence embedding models to compute distance between ideas, but does not offer an explanation: what is similar and/or different between ideas, what is being added or changed. Our most recent model extracts explicit relations between key nouns in different ideas using a combination of embedding vectors and Wordnet ontology. This ASPIRE research used large language models (LLMs) to extract named entities from input ideas, retrieve their descriptions from Wikipedia and use the descriptions to identify the relations between named entities and key nouns in different ideas. Also, we improved the identification of the correct definition for the key nouns by using LLMs. 11

17.

Environmental Engineering Project Retrieval Assistant Researcher: Ashmini Jaisingh Advisors: Dr. Krishnan Pillaipakkamnatt and Joseph Heaney III PE, Principal Walden Environmental Engineering

Large Language Models (LLMs) are advanced artificial intelligence (AI) systems trained on vast amounts of data to process and analyze human language. While powerful, standalone LLMs are prone to hallucinations, such as generating false, misleading, and/or fabricated information. A Retrieval-Augmented Generation (RAG) is an approach designed to reduce these hallucinations and improve accuracy. In a RAG pipeline, the system first retrieves relevant information from documents, augments the prompt with this grounded data, and then generates a response based on both the retrieved content and the LLM’s learned knowledge. This enhances accuracy, expands access to up-to-date information, and offers greater flexibility compared to a standalone LLM. However, RAG applications require significant preparation, including data ingestion, embedding generation, retrieval indexing, and seamless integration with the LLM. The quality of the generated output is directly tied to how effectively external data is transformed and indexed. To address this, I have developed preprocessing workflows to make non-textual formats usable within a RAG system. So far, I have implemented conversions for a range of file types– including CAD files, PowerPoint presentations, Word documents, Excel sheets, PDFs, and images– through methods like text extraction, image-to-text extraction (OCR), and structured format transformation. These workflows enable diverse datasets to be indexed and retrieved in a RAG application. I have built two working prototypes for this ongoing project: one running in a terminal environment and another through Open WebUI. 18.

FPGA Implementation of Combinatorial Algorithms Researcher: Joseph Capone Advisor: Dr. Simon Shamoun

Field Programmable Gate Arrays (FPGAs) are versatile and reconfigurable hardware platforms that can support a wide range of computational tasks by allowing algorithms and data structures to be directly hard-coded into the hardware. This project investigated how FPGAs can be effectively configured to execute combinatorial algorithms, which focus on optimizing or finding feasible solutions by maximizing or minimizing objective functions. While theoretical properties of combinatorial algorithms are well established, their practical deployment on hardware like FPGAs presents unique challenges, particularly due to hardware constraints and the complexity of real-world implementations. Emphasis is placed on graph algorithms, a crucial subset of combinatorial algorithms that model relationships between entities and often require dynamically varying data structures. Since graph structures differ for each problem instance, traditional hard-coding approaches are inadequate. The project investigated hardware acceleration of discrete algorithms by implementing an adjacency matrix on an FPGA and using a graph algorithm on it. This shows potential for the use of hardware acceleration of discrete algorithms beyond what is possible with traditional software approaches. 12

19.

Bird Migration Modeling and Radar Placement Optimization Researcher: Ajay Shekar Advisor: Dr. Simon Shamoun

This project focuses on the development of an interactive tool to analyze bird migration patterns and determine optimal radar placement for monitoring bird activity. Using realworld avian flight data (latitude, longitude, timestamp, altitude, and velocity), the system processes and cleans raw CSV datasets and visualizes the data using interactive Streamlit and Folium maps. The main part of the application solves for the location with the highest observed bird density over time. The algorithm separates the map into a grid and uses spatial filtering combined with a configurable detection radius to evaluate bird counts in each cell. The radius with the highest count is proposed as the ideal radar placement location. The final output includes a JSON file with the best location coordinates and bird count, with an interactive map overlay to visualize the results inside the map. By combining data science with geospatial analysis, the tool offers a practical approach to ecological monitoring and resource optimization for researchers that are pursuing this bird migration and may have more raw CSV datasets.

20.

Mitigating Excessive vCPU Spinning in Confidential VMs Without Hypervisor Support Researcher: Nickolaos Hatzigeorgiou Advisor: Dr. Jianchen Shan

Objective: To design and implement a lightweight, guest-only solution to eliminate wasteful vCPU spinning in Confidential VMs, where hypervisor cooperation and paravirtualization are not feasible. Background and Motivation: In modern Confidential VM environments, excessive spinning on locks remains a major inefficiency, especially when the lock holder’s vCPU is preempted. Techniques like adaptive spinning and deferred preemption—effective in host kernels—fail in guest kernels due to the lack of visibility into vCPU states and inability to influence scheduling. This problem is exacerbated in multi-cloud and untrusted host scenarios, where guest VMs cannot rely on hypervisor support. Key Idea: We propose a novel mechanism that enables adaptive spinning and approximates deferred preemption entirely within the guest kernel. Our approach uses a lightweight method to detect if a vCPU is preempted—without paravirtualization or hypervisor involvement. This allows the lock waiter to skip spinning when the lock holder is not actively executing. Additionally, we dynamically migrate lock-holding threads away from vCPUs that are likely to be preempted, approximating deferred preemption.

21.

Topological Pathways To Hallucination Detection: A Literature Review. Researchers: Hamza Virk & Manmohit Nijjer Advisor: Dr. Marco Romanelli

Large Language Models (LLMs) exhibit remarkable generative capabilities but remain vulnerable to hallucinations—producing outputs that are factually incorrect or structurally incoherent. Recent work has explored the use of topological data analysis (TDA) to interpret internal model behavior, including the application of persistent homology to attention structures. Building on this emerging literature, we conduct a focused review of topological approaches to hallucination detection, with particular attention to the mathematical foundations and interpretability benefits of persistent homology. We examine prior frameworks such as TOHA and its successors, which extend homological analysis beyond zeroth-order features, and assess their theoretical guarantees, dimensional coverage, and practical limitations. Through this review, we identify key points for future methods, including information completeness, additive mismatch sensitivity, and Lipschitz stability of topological summaries. Our synthesis clarifies the landscape of topological reliability metrics and motivates the development of principled, model-internal detectors for hallucination in LLMs. In which doing so will allow us to begin working towards achieving a new discovery in the future. 22.

Critical Vulnerabilities of Machine Learning Classifiers Susceptible to Adversarial Doppelganger Attacks Researcher: Samantha Nadler Advisor: Dr. Gerda Kamberova

Adversarial doppelganger (AD) attacks are a phenomenon in machine learning (ML) training that occur when an ML classifier assigns different labels to inputs that are indistinguishable to humans. These attacks are performed to intentionally fool the model and exploit the vulnerabilities of high-performance classifiers that can easily distinguish two distinct images as the same or different but fail to consistently correctly label “similar” images. Efforts have been made in ML research to mitigate this phenomenon through robust training and salience-based training, but the accuracy-robustness tradeoff limits such training to models with mid-range accuracy and robustness. The chosen AD attack for this research was a hybrid approach, combining Fast Gradient Sign Method (FGSM) with psychophysical concepts such as Weber’s Law and the just-noticeable difference to create attacks in which adjacent images in a chain of attacks are visually indistinguishable for human viewers. Adversarial chains were created with handwritten digit images from the MNIST dataset and then colorized using the Viridis colormap to observe label changes and doppelgangers on a built-from-scratch 12-layer convolutional neural network (CNN). Over 70% of roughly 10,000 chains created have at least one label change (and thus, one doppelganger), with some chains having as many as four label changes. The base CNN model had an initial accuracy exceeding 98% when fitted on the original MNIST train data and evaluated on its test data. After creating a doppelganger augmented dataset, using the median images of each adversarial chain, roughly 35% of perturbed images retained their original label. Recompiling and refitting the model on the augmented data reduced the model’s accuracy by at least 3% when re-evaluated on the original test data. Future steps include replicating this process on two selected state-of-the-art image classifiers whose architectures and behaviors deviate from a standard CNN model: the Branching/Merging CNN and Efficient-CapsNet models. 14

23.

Can small VLMs Drive? An experimental evaluation of small Vision Language Models for Autonomous Driving. Researcher: Samson Mathew Advisor: Dr. Zonghua Gu

Foundation Models (FMs) are Deep Learning models that are pre-trained on large unlabeled datasets through self-supervision and then fine-tuned for different downstream tasks, including. Large Language Models (LLMs) and Vision Language Models (VLMs), etc. FMs bring generalist knowledge gained during pretraining to enable human-level intelligence for Autonomous Driving (with or without fine-tuning). Especially, recent advancements in VLMs have sparked interest in their use for AD, with applications in perception and scene understanding, planning and decision-making, trajectory prediction, UI & personalization, and end-to-end driving. However, their practical adoption is hindered by challenges in deployment and real-time inference of large VLMs on in-vehicle computing platforms with limited hardware resources, as they are quite resourcedemanding due to their sheer size. Recently, researchers developed DriveBench, a large-scale benchmark dataset designed to evaluate VLM reliability for Autonomous Driving, which included the evaluation of 12 large VLMs on the dataset. In this study, we comprehensively evaluate a number of wellknown and recently developed small VLMs on the dataset, and attempt to answer the question “Can small VLMs Drive?” While the general conclusion is “no”, this work aims to provide a useful reference for researchers and practitioners in evaluating the tradeoffs between performance and computation efficiency for selecting appropriate VLMs for AD research and practice. 24.

AI Application within RCEs Researcher: Manuel Lizarazo Advisor: Dr. Oren Segal

Though our data storage systems seem to continue to scale in size with no slow-down in sight, there will always appear to be places where these solutions simply don’t have applicability. In these Resource-Constrained Environments (RCE), it’s essential to be mindful of total resources and how we choose to allocate them. In a world where AI seems to be growing in size just as quickly, at first glance, it would seem that the two, AI in RCEs, are incompatible. I propose a solution to this compatibility problem in the form of a novel benchmarking software that would allow a user to input their LLM model and have it automatically, through distillation and layer compression techniques, output a competitively performing RCE-optimized model. A model which would decrease the total resources required while preserving most of the original input model’s accuracy.

25.

Optimizing Question Answering Accuracy with Retrieval-Augmented Generation Researcher: Stella Sourelos Advisor: Dr. Oren Segal

This project explores the integration of a Retrieval-Augmented Generation (RAG) pipeline with a GPT-2-based Large Language Model (LLM) to improve response accuracy and contextual relevance for question-answering tasks. RAG combines the generative capabilities of LLMs with dynamic retrieval from an external knowledge base, mitigating hallucinations and enhancing factual precision. Two modes were evaluated: RAG-enabled, where relevant context from a SQLite knowledge database is retrieved and incorporated into model prompts, and RAG-disabled, where generation relies solely on the pre-trained model. Performance was assessed using Bilingual Evaluation Understudy (BLEU) scores, semantic similarity, and an overall blended metric. The results consistently demonstrated that the RAG pipeline significantly improves the quality, relevance, and factual grounding of generated responses compared to a baseline without retrieval. The ultimate goal of this research is to create a hardware-aware RAG framework capable of delivering accurate, context-rich responses in environments where computational resources are limited.

26.

Lightweight Reliable Token Streaming with UDP: An Open-Source Implementation Inspired by Eloquent Researcher: Sajan Kavuri Advisor: Dr. Oren Segal

Large Language Models (LLMs) and other real-time applications depend on fast, reliable token delivery, but existing transport protocols are often too heavy for ResourceConstrained Environments (RCE). The recently proposed Eloquent protocol introduced a lightweight transport approach, but its implementation is not open-source, and its evaluation has so far been limited to controlled research settings. This work introduces an open-source partial-prototype inspired by Eloquent, implemented over a reliable User Datagram Protocol (UDP) layer for lightweight systems. The design ensures ordered token delivery with acknowledgments and retransmissions, while avoiding the overhead of Transmission Control Protocol (TCP). A server serializes and streams tokens, and a client reconstructs and verifies them. Current testing confirms the protocol functions correctly, with both server and client able to operate and exit gracefully, enabling automation without manual intervention. Our work shows that reliable token streaming can be achieved with a minimal and transparent protocol suited for RCE. 16

LAST PAGE